Generalist AI unveils GEN-1.5 for learning robot tasks from one demo
Robotics startup Generalist AI has introduced GEN-1.5, an AI model designed to let robots perform unfamiliar tasks after watching a single demonstration. The demonstration can last from three to 12 seconds and is placed in the model’s context window as a “physical prompt,” functioning as short-term memory. The company says the robot then carries out the task without additional training.
In ten tests, including opening a jar and taking money from a wallet, Generalist reported an average success rate of 59 percent. Ten training steps using five minutes of data raised the reported rate to 83 percent. GEN-1.5 can also combine two prompts into longer action sequences, accept demonstrations from simulations and partially copy human hand movements.
Generalist says these capabilities emerged during more than eight months of pretraining on interaction data and were not directly taught. Similar in-context learning has been demonstrated by other research groups, but the company says GEN-1.5 applies it to a broader range of tasks. The examples remain short and simple, and the results have not been independently verified.
Written by Sakhanda Wire AI from reporting by The Decoder. Facts, quotes and full context are in the original article below.